NPU 시스템 확장을 위한 데이터 전송 최적화

Abstract

As the size and complexity of artificial intelligence models increase, the need for distributed computing to efficiently handle large-scale computation and data transfer is becoming increasingly important. However, in a hardware environment that does not support peer-to-peer communication, data transfer between accelerators is limited, which makes it difficult to implement distributed computing. In this paper, we analyse the performance of data transfer between multiple neural processing units (NPUs) based on Peripheral Component Interconnect Express (PCIe) using the Neubla Antara NPU and propose two transfer methods to solve this problem. The first is a Base Address Register (BAR)-based transmission method, in which the central processing unit (CPU) copies and transmits data between NPUs. The second is a CPU-direct memory access (DMA) cooperative transmission method, in which data transmission between the device and the host is handled by DMA, and the host transfers data in stages by the CPU. The results of the experiment showed that the BAR-based transmission recorded a speed of 24.30 MB/s, while the CPU-DMA cooperative-based transmission recorded a speed of 1988.28 MB/s. This study experimentally compares and analyses the performance of the two transmission methods and presents a solution that can support data transmission in an environment where peer-to-peer communication is not possible.